1 /* -*- mode: c; c-basic-offset: 8; -*-
2 * vim: noexpandtab sw=8 ts=8 sts=0:
6 * underlying calls for lock conversion
8 * Copyright (C) 2004 Oracle. All rights reserved.
10 * This program is free software; you can redistribute it and/or
11 * modify it under the terms of the GNU General Public
12 * License as published by the Free Software Foundation; either
13 * version 2 of the License, or (at your option) any later version.
15 * This program is distributed in the hope that it will be useful,
16 * but WITHOUT ANY WARRANTY; without even the implied warranty of
17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
18 * General Public License for more details.
20 * You should have received a copy of the GNU General Public
21 * License along with this program; if not, write to the
22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
23 * Boston, MA 021110-1307, USA.
28 #include <linux/module.h>
30 #include <linux/types.h>
31 #include <linux/slab.h>
32 #include <linux/highmem.h>
33 #include <linux/utsname.h>
34 #include <linux/init.h>
35 #include <linux/sysctl.h>
36 #include <linux/random.h>
37 #include <linux/blkdev.h>
38 #include <linux/socket.h>
39 #include <linux/inet.h>
40 #include <linux/spinlock.h>
43 #include "cluster/heartbeat.h"
44 #include "cluster/nodemanager.h"
45 #include "cluster/tcp.h"
48 #include "dlmcommon.h"
50 #include "dlmconvert.h"
52 #define MLOG_MASK_PREFIX ML_DLM
53 #include "cluster/masklog.h"
55 /* NOTE: __dlmconvert_master is the only function in here that
56 * needs a spinlock held on entry (res->spinlock) and it is the
57 * only one that holds a lock on exit (res->spinlock).
58 * All other functions in here need no locks and drop all of
59 * the locks that they acquire. */
60 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
61 struct dlm_lock_resource *res,
62 struct dlm_lock *lock, int flags,
63 int type, int *call_ast,
65 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
66 struct dlm_lock_resource *res,
67 struct dlm_lock *lock, int flags, int type);
70 * this is only called directly by dlmlock(), and only when the
71 * local node is the owner of the lockres
74 * taken: takes and drops res->spinlock
76 * returns: see __dlmconvert_master
78 enum dlm_status dlmconvert_master(struct dlm_ctxt *dlm,
79 struct dlm_lock_resource *res,
80 struct dlm_lock *lock, int flags, int type)
82 int call_ast = 0, kick_thread = 0;
83 enum dlm_status status;
85 spin_lock(&res->spinlock);
86 /* we are not in a network handler, this is fine */
87 __dlm_wait_on_lockres(res);
88 __dlm_lockres_reserve_ast(res);
89 res->state |= DLM_LOCK_RES_IN_PROGRESS;
91 status = __dlmconvert_master(dlm, res, lock, flags, type,
92 &call_ast, &kick_thread);
94 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
95 spin_unlock(&res->spinlock);
97 if (status != DLM_NORMAL && status != DLM_NOTQUEUED)
100 /* either queue the ast or release it */
102 dlm_queue_ast(dlm, lock);
104 dlm_lockres_release_ast(dlm, res);
107 dlm_kick_thread(dlm, res);
112 /* performs lock conversion at the lockres master site
114 * caller needs: res->spinlock
115 * taken: takes and drops lock->spinlock
116 * held on exit: res->spinlock
117 * returns: DLM_NORMAL, DLM_NOTQUEUED, DLM_DENIED
118 * call_ast: whether ast should be called for this lock
119 * kick_thread: whether dlm_kick_thread should be called
121 static enum dlm_status __dlmconvert_master(struct dlm_ctxt *dlm,
122 struct dlm_lock_resource *res,
123 struct dlm_lock *lock, int flags,
124 int type, int *call_ast,
127 enum dlm_status status = DLM_NORMAL;
128 struct list_head *iter;
129 struct dlm_lock *tmplock=NULL;
131 assert_spin_locked(&res->spinlock);
133 mlog_entry("type=%d, convert_type=%d, new convert_type=%d\n",
134 lock->ml.type, lock->ml.convert_type, type);
136 spin_lock(&lock->spinlock);
138 /* already converting? */
139 if (lock->ml.convert_type != LKM_IVMODE) {
140 mlog(ML_ERROR, "attempted to convert a lock with a lock "
141 "conversion pending\n");
146 /* must be on grant queue to convert */
147 if (!dlm_lock_on_list(&res->granted, lock)) {
148 mlog(ML_ERROR, "attempted to convert a lock not on grant "
154 if (flags & LKM_VALBLK) {
155 switch (lock->ml.type) {
157 /* EX + LKM_VALBLK + convert == set lvb */
158 mlog(0, "will set lvb: converting %s->%s\n",
159 dlm_lock_mode_name(lock->ml.type),
160 dlm_lock_mode_name(type));
161 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
165 /* refetch if new level is not NL */
166 if (type > LKM_NLMODE) {
167 mlog(0, "will fetch new value into "
168 "lvb: converting %s->%s\n",
169 dlm_lock_mode_name(lock->ml.type),
170 dlm_lock_mode_name(type));
171 lock->lksb->flags |= DLM_LKSB_GET_LVB;
173 mlog(0, "will NOT fetch new value "
174 "into lvb: converting %s->%s\n",
175 dlm_lock_mode_name(lock->ml.type),
176 dlm_lock_mode_name(type));
177 flags &= ~(LKM_VALBLK);
184 /* in-place downconvert? */
185 if (type <= lock->ml.type)
188 /* upconvert from here on */
190 list_for_each(iter, &res->granted) {
191 tmplock = list_entry(iter, struct dlm_lock, list);
194 if (!dlm_lock_compatible(tmplock->ml.type, type))
198 list_for_each(iter, &res->converting) {
199 tmplock = list_entry(iter, struct dlm_lock, list);
200 if (!dlm_lock_compatible(tmplock->ml.type, type))
202 /* existing conversion requests take precedence */
203 if (!dlm_lock_compatible(tmplock->ml.convert_type, type))
207 /* fall thru to grant */
210 mlog(0, "res %.*s, granting %s lock\n", res->lockname.len,
211 res->lockname.name, dlm_lock_mode_name(type));
212 /* immediately grant the new lock type */
213 lock->lksb->status = DLM_NORMAL;
214 if (lock->ml.node == dlm->node_num)
215 mlog(0, "doing in-place convert for nonlocal lock\n");
216 lock->ml.type = type;
217 if (lock->lksb->flags & DLM_LKSB_PUT_LVB)
218 memcpy(res->lvb, lock->lksb->lvb, DLM_LVB_LEN);
225 if (flags & LKM_NOQUEUE) {
226 mlog(0, "failed to convert NOQUEUE lock %.*s from "
227 "%d to %d...\n", res->lockname.len, res->lockname.name,
228 lock->ml.type, type);
229 status = DLM_NOTQUEUED;
232 mlog(0, "res %.*s, queueing...\n", res->lockname.len,
235 lock->ml.convert_type = type;
236 /* do not alter lock refcount. switching lists. */
237 list_move_tail(&lock->list, &res->converting);
240 spin_unlock(&lock->spinlock);
241 if (status == DLM_DENIED) {
242 __dlm_print_one_lock_resource(res);
244 if (status == DLM_NORMAL)
249 void dlm_revert_pending_convert(struct dlm_lock_resource *res,
250 struct dlm_lock *lock)
252 /* do not alter lock refcount. switching lists. */
253 list_move_tail(&lock->list, &res->granted);
254 lock->ml.convert_type = LKM_IVMODE;
255 lock->lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
258 /* messages the master site to do lock conversion
261 * taken: takes and drops res->spinlock, uses DLM_LOCK_RES_IN_PROGRESS
263 * returns: DLM_NORMAL, DLM_RECOVERING, status from remote node
265 enum dlm_status dlmconvert_remote(struct dlm_ctxt *dlm,
266 struct dlm_lock_resource *res,
267 struct dlm_lock *lock, int flags, int type)
269 enum dlm_status status;
271 mlog(0, "type=%d, convert_type=%d, busy=%d\n", lock->ml.type,
272 lock->ml.convert_type, res->state & DLM_LOCK_RES_IN_PROGRESS);
274 spin_lock(&res->spinlock);
275 if (res->state & DLM_LOCK_RES_RECOVERING) {
276 mlog(0, "bailing out early since res is RECOVERING "
277 "on secondary queue\n");
278 /* __dlm_print_one_lock_resource(res); */
279 status = DLM_RECOVERING;
282 /* will exit this call with spinlock held */
283 __dlm_wait_on_lockres(res);
285 if (lock->ml.convert_type != LKM_IVMODE) {
286 __dlm_print_one_lock_resource(res);
287 mlog(ML_ERROR, "converting a remote lock that is already "
288 "converting! (cookie=%u:%llu, conv=%d)\n",
289 dlm_get_lock_cookie_node(be64_to_cpu(lock->ml.cookie)),
290 dlm_get_lock_cookie_seq(be64_to_cpu(lock->ml.cookie)),
291 lock->ml.convert_type);
295 res->state |= DLM_LOCK_RES_IN_PROGRESS;
296 /* move lock to local convert queue */
297 /* do not alter lock refcount. switching lists. */
298 list_move_tail(&lock->list, &res->converting);
299 lock->convert_pending = 1;
300 lock->ml.convert_type = type;
302 if (flags & LKM_VALBLK) {
303 if (lock->ml.type == LKM_EXMODE) {
304 flags |= LKM_PUT_LVB;
305 lock->lksb->flags |= DLM_LKSB_PUT_LVB;
307 if (lock->ml.convert_type == LKM_NLMODE)
308 flags &= ~LKM_VALBLK;
310 flags |= LKM_GET_LVB;
311 lock->lksb->flags |= DLM_LKSB_GET_LVB;
315 spin_unlock(&res->spinlock);
317 /* no locks held here.
318 * need to wait for a reply as to whether it got queued or not. */
319 status = dlm_send_remote_convert_request(dlm, res, lock, flags, type);
321 spin_lock(&res->spinlock);
322 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
323 lock->convert_pending = 0;
324 /* if it failed, move it back to granted queue */
325 if (status != DLM_NORMAL) {
326 if (status != DLM_NOTQUEUED)
328 dlm_revert_pending_convert(res, lock);
331 spin_unlock(&res->spinlock);
333 /* TODO: should this be a wake_one? */
334 /* wake up any IN_PROGRESS waiters */
340 /* sends DLM_CONVERT_LOCK_MSG to master site
345 * returns: DLM_NOLOCKMGR, status from remote node
347 static enum dlm_status dlm_send_remote_convert_request(struct dlm_ctxt *dlm,
348 struct dlm_lock_resource *res,
349 struct dlm_lock *lock, int flags, int type)
351 struct dlm_convert_lock convert;
358 mlog_entry("%.*s\n", res->lockname.len, res->lockname.name);
360 memset(&convert, 0, sizeof(struct dlm_convert_lock));
361 convert.node_idx = dlm->node_num;
362 convert.requested_type = type;
363 convert.cookie = lock->ml.cookie;
364 convert.namelen = res->lockname.len;
365 convert.flags = cpu_to_be32(flags);
366 memcpy(convert.name, res->lockname.name, convert.namelen);
368 vec[0].iov_len = sizeof(struct dlm_convert_lock);
369 vec[0].iov_base = &convert;
371 if (flags & LKM_PUT_LVB) {
372 /* extra data to send if we are updating lvb */
373 vec[1].iov_len = DLM_LVB_LEN;
374 vec[1].iov_base = lock->lksb->lvb;
378 tmpret = o2net_send_message_vec(DLM_CONVERT_LOCK_MSG, dlm->key,
379 vec, veclen, res->owner, &status);
381 // successfully sent and received
382 ret = status; // this is already a dlm_status
383 if (ret == DLM_RECOVERING) {
384 mlog(0, "node %u returned DLM_RECOVERING from convert "
385 "message!\n", res->owner);
386 } else if (ret == DLM_MIGRATING) {
387 mlog(0, "node %u returned DLM_MIGRATING from convert "
388 "message!\n", res->owner);
389 } else if (ret == DLM_FORWARD) {
390 mlog(0, "node %u returned DLM_FORWARD from convert "
391 "message!\n", res->owner);
392 } else if (ret != DLM_NORMAL && ret != DLM_NOTQUEUED)
396 if (dlm_is_host_down(tmpret)) {
397 /* instead of logging the same network error over
398 * and over, sleep here and wait for the heartbeat
399 * to notice the node is dead. times out after 5s. */
400 dlm_wait_for_node_death(dlm, res->owner,
401 DLM_NODE_DEATH_WAIT_MAX);
402 ret = DLM_RECOVERING;
403 mlog(0, "node %u died so returning DLM_RECOVERING "
404 "from convert message!\n", res->owner);
406 ret = dlm_err_to_dlm_status(tmpret);
413 /* handler for DLM_CONVERT_LOCK_MSG on master site
416 * taken: takes and drop res->spinlock
418 * returns: DLM_NORMAL, DLM_IVLOCKID, DLM_BADARGS,
419 * status from __dlmconvert_master
421 int dlm_convert_lock_handler(struct o2net_msg *msg, u32 len, void *data,
424 struct dlm_ctxt *dlm = data;
425 struct dlm_convert_lock *cnv = (struct dlm_convert_lock *)msg->buf;
426 struct dlm_lock_resource *res = NULL;
427 struct list_head *iter;
428 struct dlm_lock *lock = NULL;
429 struct dlm_lockstatus *lksb;
430 enum dlm_status status = DLM_NORMAL;
432 int call_ast = 0, kick_thread = 0, ast_reserved = 0, wake = 0;
434 if (!dlm_grab(dlm)) {
435 dlm_error(DLM_REJECTED);
439 mlog_bug_on_msg(!dlm_domain_fully_joined(dlm),
440 "Domain %s not fully joined!\n", dlm->name);
442 if (cnv->namelen > DLM_LOCKID_NAME_MAX) {
443 status = DLM_IVBUFLEN;
448 flags = be32_to_cpu(cnv->flags);
450 if ((flags & (LKM_PUT_LVB|LKM_GET_LVB)) ==
451 (LKM_PUT_LVB|LKM_GET_LVB)) {
452 mlog(ML_ERROR, "both PUT and GET lvb specified\n");
453 status = DLM_BADARGS;
457 mlog(0, "lvb: %s\n", flags & LKM_PUT_LVB ? "put lvb" :
458 (flags & LKM_GET_LVB ? "get lvb" : "none"));
460 status = DLM_IVLOCKID;
461 res = dlm_lookup_lockres(dlm, cnv->name, cnv->namelen);
467 spin_lock(&res->spinlock);
468 status = __dlm_lockres_state_to_status(res);
469 if (status != DLM_NORMAL) {
470 spin_unlock(&res->spinlock);
474 list_for_each(iter, &res->granted) {
475 lock = list_entry(iter, struct dlm_lock, list);
476 if (lock->ml.cookie == cnv->cookie &&
477 lock->ml.node == cnv->node_idx) {
483 spin_unlock(&res->spinlock);
485 status = DLM_IVLOCKID;
486 mlog(ML_ERROR, "did not find lock to convert on grant queue! "
488 dlm_get_lock_cookie_node(be64_to_cpu(cnv->cookie)),
489 dlm_get_lock_cookie_seq(be64_to_cpu(cnv->cookie)));
490 __dlm_print_one_lock_resource(res);
497 /* see if caller needed to get/put lvb */
498 if (flags & LKM_PUT_LVB) {
499 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
500 lksb->flags |= DLM_LKSB_PUT_LVB;
501 memcpy(&lksb->lvb[0], &cnv->lvb[0], DLM_LVB_LEN);
502 } else if (flags & LKM_GET_LVB) {
503 BUG_ON(lksb->flags & (DLM_LKSB_PUT_LVB|DLM_LKSB_GET_LVB));
504 lksb->flags |= DLM_LKSB_GET_LVB;
507 spin_lock(&res->spinlock);
508 status = __dlm_lockres_state_to_status(res);
509 if (status == DLM_NORMAL) {
510 __dlm_lockres_reserve_ast(res);
512 res->state |= DLM_LOCK_RES_IN_PROGRESS;
513 status = __dlmconvert_master(dlm, res, lock, flags,
515 &call_ast, &kick_thread);
516 res->state &= ~DLM_LOCK_RES_IN_PROGRESS;
519 spin_unlock(&res->spinlock);
523 if (status != DLM_NORMAL) {
524 if (status != DLM_NOTQUEUED)
526 lksb->flags &= ~(DLM_LKSB_GET_LVB|DLM_LKSB_PUT_LVB);
533 /* either queue the ast or release it, if reserved */
535 dlm_queue_ast(dlm, lock);
536 else if (ast_reserved)
537 dlm_lockres_release_ast(dlm, res);
540 dlm_kick_thread(dlm, res);
543 dlm_lockres_put(res);